The Daemon

Status: Tier 2. Repo: fermi-daemon. A set of small processes on your Mac that turn the Worker's task queue into real work done by a real coding harness.

The core idea

The Worker cannot run Claude Code. Your Mac can. The daemon closes that gap pull-only: the Mac calls out to the Worker, claims tasks, runs the harness locally, and completes them. Without MacOSMCP, nothing on the internet can push into your Mac — the daemon holds an outbound bearer token, no inbound port.

What runs (who owns what)

Process	File	Owns
Poller	poll.sh	lane-based task drain: claims from its queue, spawns a fresh harness per task, completes with the result
Warm worker	warm-worker.mjs	latency: keeps one harness pre-booted so the first token of a reply is seconds, not a cold boot; drains main
Channel bridges	wa-bridge, dc-bridge (in fermi-mcp packages/)	sockets a Worker can't hold (WhatsApp Web); relay to Worker webhooks
Broker executor	tools/broker-executor.mjs	the only process anywhere that sees decrypted cookies; drives Playwright for fleet browser ops — see Session broker
Session capture	tools/capture-session.mjs	logs into a site headfully once, ships storageState to /admin/session/capture
Box runner	box-runner.mjs	not run on the Mac — this is the file neutrinos boot; it lives here so it can be SHA-pinned from one repo

With vs. without MacOSMCP

	Daemon only (Tier 2)	+ MacOSMCP (Tier 3)
Direction	Mac pulls tasks	Worker can also push tool calls in
Latency	queue poll (seconds)	live round-trip
Surface	what your task prompt allows the harness	25 mac_* tools for any connected host
Inbound exposure	none	Cloudflare Tunnel + MACOS_MCP_TOKEN
Typical use	"fix issue #12 and open a PR" from Telegram	"log into the vendor portal and screenshot the invoice" from your phone

They compose: the reference deployment runs both. Tier 2 handles queued work; Tier 3 handles interactive control. The daemon's harness itself connects to Fermi over MCP, so queued tasks still get memory, skills, secrets, and channels.

Phone calls from a chat

A lane lives for minutes; a call can last hours. So a call requested from a channel is split in two: one lane confirms the details with you, starts the call with phone_call_start, and finishes; the call then runs inside the Worker; and when it ends the Worker enqueues a result task that a fresh lane relays to your chat. The drain prompt requires the lane to state who it will call and what it will say, and to wait for your yes before dialing. Details: Phone calls.

Setup

git clone https://github.com/abel30567/fermi-daemon && cd fermi-daemon
cp .env.example .env    # FERMI_URL, FERMI_BEARER_TOKEN, lane/queue names
bun install             # playwright, for the broker executor
./poll.sh               # or install the launchd plists for boot persistence
node warm-worker.mjs    # optional but worth it

What success looks like

• [ ] Message your Telegram bot; a reply arrives produced on your Mac
• [ ] task_list(status: done) shows the task claimed by your daemon's identity
• [ ] Kill the daemon mid-task: the lease expires and the task becomes claimable again (no lost work)

Failure modes

Symptom	Cause	Fix
Tasks pile up pending	daemon not running / wrong queue name	check lane config vs. enqueue queue
Replies from "root" or wrong git author	harness env leaks through	daemon exports GIT_AUTHOR_*/GIT_COMMITTER_*; see Committing as root
Fleet tasks stolen by your Mac	you're on a pre-Sep-2026 worker	update: fleet queues are excluded from the channel drain
Two broker executors fighting	you ran it on two Macs	exactly one owns the browser profiles; kill the rest